Prompt injection prevention is best described by which practice?

• A. Evaluating and sanitizing data input.
• B. Expanding the model architecture.
• C. Disabling all prompts entirely.
• D. Storing prompts in plaintext.

Answer: (A)

Preventing prompt injection centers on controlling what the model sees by validating and sanitizing input. Attackers can weave instructions or hidden prompts into user-provided data to steer the model, reveal sensitive information, or bypass safety checks. By evaluating input and filtering or neutralizing dangerous content before it reaches the model, you remove or diminish the influence of those embedded directives. This involves removing or escaping special tokens, applying safe encoding, enforcing allowlists of acceptable constructs, and limiting input structure or length so malicious patterns can’t commandeer the model’s behavior. In practice, input sanitization acts as a first line of defense that reduces the chance that a user’s text will override the intended prompts or system rules.

Expanding the model architecture doesn’t solve the problem at the input stage, and disabling prompts entirely would render the tool impractical. Storing prompts in plaintext increases exposure risk without addressing how harmful input could manipulate the model.